Epidemic mixing data prediction method based on deep learning

Abstract

The invention relates to an epidemic mixing data prediction method based on deep learning, and the method comprises the following steps: obtaining historical epidemic situation data, real-time monitoring data and demographic data, and obtaining mixing data through information fusion; determining a sliding window according to the optimal lag order of the frequency mixing data and the sampling time interval, and determining the data length of the data of each sampling frequency in the window; obtaining a frequency mixing data embedding vector; adaptively extracting or expanding time sequence characteristics in a time dimension; learning attention vectors from the fusion features, and obtaining time and attribute feature context vectors; the frequency mixing data embedding vector and the time and attribute feature context vector are output through a full connection layer to obtain a predicted value; according to the invention, through processing the frequency mixing data, the accuracy and timeliness of epidemic disease prediction are improved, the deep learning model is utilized to process complex data, the prediction accuracy and timeliness of epidemic disease outbreak and propagation trend can be effectively improved, and a scientific basis is provided for public health decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of epidemic prediction, and in particular to a method for predicting epidemic mixed frequency data based on deep learning. Background Art

[0002] In the prior art, the prediction models based on statistical methods are mainly used for single-step or short-term trend prediction of infectious diseases, model prediction based on machine learning and deep learning, and various combination models of these technologies, which have shown excellent results in multiple epidemic prediction research scenarios. However, these methods still have shortcomings and need further optimization and development. First, the solution for multi-step prediction generally adopts an iterative prediction method, that is, the predicted value obtained from each single-step prediction is added to the existing sequence, and then a new single-step prediction is started with the new sequence, and the long-term prediction results are obtained by repeating multiple times. The defect of this solution is that it is impossible to obtain the prediction results of multiple time steps through a single prediction task, and the sequence generation model in the field of natural language processing provides inspiration for the realization of "one-step" multi-step prediction. Secondly, from the data point of view, the data of epidemic influencing factors usually have different frequencies, such as regional weather data is calculated on a daily basis, while the incidence rate is calculated on a weekly basis, the regional economic indicators are calculated on a monthly basis, and the population indicators are calculated on an annual basis; although there are related studies at home and abroad, there is still a lack of research on the prediction and analysis of multi-dimensional mixed frequency data of epidemics. Summary of the invention

[0003] The purpose of the present invention is to provide a method for predicting epidemic mixed frequency data based on deep learning, aiming to improve the accuracy and efficiency of epidemic trend prediction.

[0004] In order to solve the above technical problems, a method for predicting epidemic mixed frequency data based on deep learning is provided, which includes the following steps:

[0005] S1. Obtain historical epidemic data, real-time monitoring data, and demographic data, and associate data of different frequencies to obtain mixed frequency data through information fusion;

[0006] S2, determining the size of the sliding window according to the optimal lag order and sampling time interval of the mixing data, and determining the data length of the data of each sampling frequency in the window when the sliding window is initialized; constructing a city network according to the geographical distribution map of the region and its adjacent information;

[0007] S3, learn the time-dependent features of each frequency data through the LSTM network and obtain the mixed frequency data embedding vector;

[0008] S4, adaptively extracting or expanding the timing characteristics in the time dimension through a CNN network with the same number of convolution kernels, wherein the CNN network unifies the mixed data into the same data length and fuses them to obtain fusion features, so that the feature layer information of the mixed data is fused;

[0009] S5. Learning an attention vector from the fused feature through a time attribute attention mechanism, and obtaining a time and attribute feature context vector, wherein the attention mechanism adaptively mines important features in the fused feature;

[0010] S6. The mixed data embedding vector, time and attribute feature context vector are output through a fully connected layer to obtain a prediction value.

[0011] Furthermore, the temporal attribute attention mechanism is a two-dimensional attention mechanism, which extracts information from two dimensions, time and attribute, respectively, so that the temporal attribute attention mechanism generates two context-related vectors.

[0012] Furthermore, the mixed frequency data needs to be processed, and the data processing includes successively performing data preprocessing and data sequence stabilization.

[0013] Furthermore, the data preprocessing includes data cleaning, missing value processing and outlier detection.

[0014] Furthermore, the data series smoothing includes differencing and deterministic detrending.

[0015] Furthermore, the data series is tested for stationarity by ADF and KPSS. When the ADF test data series result is stationary and the KPSS test data series result is non-stationary, the time series is differentially stationary - the time series is differentially applied and the stationarity is checked again. When the ADF test result is non-stationary and the KPSS test result is stationary, the time series is trend-stationary and requires deterministic detrending and stationarity is checked again.

[0016] Furthermore, in step S2, the construction of the urban network according to the geographical distribution map of the region and its adjacent information specifically includes, firstly, according to the definition of the geographical distribution map, the regional network of the monitored area is represented as an undirected graph G = (V, E, W), wherein each node represents a region, and each edge represents the adjacency relationship between two regions; represents a set of M vertices, represents a set of N edges in the network, and the relationship between each edge can be represented by two vertices (vi, vj); W represents an M*M spatial weight matrix, that is, vertices vi and vj are adjacent, then W ijis assigned a value of 1, otherwise it is assigned a value of 0; the mixed-frequency time series data related to the epidemic in each region are synchronized to form multiple sets of homofrequency time series data, whose structure is St = {s1, s2, ..., st}; the data are combined with the regional network to construct a set of dynamic infectious disease regional network diagrams that change over time.

[0017] Furthermore, the dynamic infectious disease regional network includes spatiotemporal data processing, multi-channel fusion learning and network enhanced connection strategy.

[0018] Implementing the embodiments of the present invention will have the following beneficial effects:

[0019] 1. The method for predicting mixed epidemic data based on deep learning in this embodiment improves the accuracy and timeliness of epidemic prediction by processing mixed data. Using deep learning models to process complex data can effectively improve the accuracy and timeliness of predicting epidemic outbreaks and transmission trends, and provide a scientific basis for public health decision-making;

[0020] 2. The method for predicting epidemic mixed data based on deep learning in this embodiment can help the government and health departments formulate more effective epidemic prevention and control measures, rationally allocate medical resources, and reduce the pressure on the public health system by enhancing the scientific nature of the epidemic response strategy.

[0021] 3. The method for predicting epidemic mixed data based on deep learning in this embodiment promotes the application of artificial intelligence technology in the field of medical and health care. Through in-depth research on the epidemic prediction model based on deep learning, it can promote the widespread application of artificial intelligence technology in medical and health care fields such as epidemiology and disease prevention and control, and improve the level of intelligence of medical and health services.

[0022] 4. The method for predicting epidemic mixed data based on deep learning in this embodiment provides technical support for global epidemic early warning. In the context of globalization, epidemics often spread rapidly across borders. The epidemic prediction model based on deep learning can help the global health organization to monitor and warn of transnational infectious diseases in a timely and accurate manner, and improve the coordination and efficiency of the global response to epidemics.

[0023] 5. The method for predicting epidemic mixed frequency data based on deep learning in this embodiment promotes the cross-integration of related disciplines. The combination of deep learning technology and epidemiology can not only promote the development of public health disciplines, but also promote the cross-integration and innovation of related fields such as computer science and data science. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 A flowchart of a method for predicting epidemic mixed frequency data based on deep learning according to an embodiment of the present invention;

[0026] Figure 2 A roadmap for the time series-based epidemic mixed frequency data prediction model described in an embodiment of the present invention;

[0027] Figure 3 A roadmap for the epidemic mixed frequency data prediction model based on spatiotemporal data according to an embodiment of the present invention;

[0028] Figure 4 This is a technical roadmap for the method for predicting epidemic mixed frequency data based on deep learning described in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0030] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0032] Please refer to Figure 1-Figure 3 The embodiment of the present invention provides a method for predicting epidemic mixed frequency data based on deep learning, comprising the following steps:

[0033] S1. Obtain historical epidemic data, real-time monitoring data, and demographic data, and associate data of different frequencies to obtain mixed frequency data through information fusion;

[0034] S2. Determine the size of the sliding window according to the optimal lag order and sampling time interval of the mixed frequency data, and determine the data length of the data of each sampling frequency in the window when the sliding window is initialized; construct a city network according to the geographical distribution map of the region and its adjacent information;

[0035] S3, learn the time-dependent features of each frequency data through the LSTM network and obtain the mixed frequency data embedding vector;

[0036] S4, adaptively extracting or expanding the time series characteristics in the time dimension through a CNN network with the same number of convolution kernels, wherein the CNN network unifies the mixed data into the same data length and fuses them to obtain fused features, so that the feature layer information of the mixed data is fused;

[0037] S5. Learn the attention vector from the fused features through the time attribute attention mechanism, and obtain the time and attribute feature context vector, wherein the attention mechanism adaptively mines the important features in the fused features;

[0038] S6. The mixed data embedding vector, time and attribute feature context vector are output through the fully connected layer to obtain the predicted value. In specific applications, in most cases, mixed data is a kind of time series data. The accuracy of epidemic prediction depends on the quality and magnitude of the data to a large extent. The present invention will collect multiple data sources including historical epidemic data, real-time monitoring data, demographic data, etc. The present invention regards mixed data as a kind of multi-source time series data, and connects data of different frequencies through the strategy of information fusion, which is helpful to learn the nonlinear relationship between data of different frequencies in mixed data, which provides a new design scheme for processing mixed data. First, the size of the sliding window is determined according to the optimal lag order and sampling time interval of the target sequence, and the data length of the data of each sampling frequency in the window is determined when the sliding window is initialized. Then, a group of LSTM networks are used to learn each The time-dependent features of the frequency data are extracted and the embedding vector of the mixed data is obtained; then, a group of CNN networks with the same number of convolution kernels are used to adaptively extract or expand the time series characteristics in the time dimension. CNN unifies the mixed data into the same data length and fuses the extracted features, realizing the feature layer information fusion of the mixed data; finally, the time attribute attention mechanism learns the attention vector from the fused features and obtains the time and attribute feature context vector; the attention mechanism can adaptively mine the important features in the fused features; the final prediction value is obtained by the mixed data embedding vector, the time and attribute feature context vector through the fully connected layer output; the whole model operates under the nonlinear activation function and can effectively fit the nonlinear characteristics of the mixed data.

[0039] In one possible implementation, the temporal attribute attention mechanism is a two-dimensional attention mechanism that extracts information from the two dimensions of time and attribute, so that the temporal attribute attention mechanism generates two context-related vectors. In specific applications, the context-related vector Vt of the traditional attention mechanism is generated by weighting and calculating the output of LSTM. This setting is more reasonable in natural language processing, but it ignores the importance relationship between variables in time series analysis. The temporal attribute attention mechanism in this article is a two-dimensional attention mechanism that aims to discover important features of the temporal and attribute dimensions from the fused features; the temporal attribute attention mechanism automatically identifies the key time series from the fusion matrix t; the temporal attribute attention mechanism is a two-dimensional attention mechanism that extracts information from the two dimensions of time and attribute, so the temporal attribute attention mechanism generates two context-related vectors, which enables more information to participate in the final prediction task.

[0040] In one possible implementation, the mixed frequency data needs to be processed, and the data processing includes data preprocessing and data series stabilization in sequence; data preprocessing includes data cleaning, missing value processing and outlier detection; data series stabilization includes difference and deterministic detrending. In specific applications, the accuracy of epidemic prediction depends largely on the quality and magnitude of the data. The present invention will collect a variety of data sources including historical epidemic data, real-time monitoring data, demographic data, etc. The first step of this article is to collect a wide range of historical epidemic data, which include but are not limited to the incidence, transmission speed, geographical distribution and details of the affected population; in addition, environmental factors (such as climate change, pollution level) and socioeconomic factors (such as population density, medical resource allocation) are also taken into consideration; first, due to the complexity of multidimensional mixed frequency data, the collected data needs to undergo a rigorous cleaning process to remove errors and inconsistencies; then, the data is processed by normalization and encoding methods, and a standardized process is established to process mixed frequency data, including data cleaning, missing value processing, outlier detection and processing, etc.; ensure the quality and consistency of input data, and lay the foundation for the training of deep learning models.

[0041] In one possible implementation, the data series is tested for stationarity by ADF and KPSS. When the ADF test data series result is stationary, and the KPSS test data series result is non-stationary, the time series is differentially stationary - the time series is differentially stationary, and the stationarity is checked again; when the ADF test result is non-stationary, and the KPSS test result is stationary, the time series is trend-stationary, and deterministic detrending is required, and the stationarity is checked again. In specific applications, the data structure of time series data is different from that of traditional statistical data; the biggest difference is that traditional random variables can obtain multiple observations (such as the number of dice points, which can be repeatedly thrown to obtain multiple observations, ignoring the time difference); while in time series data, each random variable has only one observation. In this way, only one observation can be obtained for each distribution, which is too small to study the properties of the distribution; however, through stationarity, the intrinsic correlation between the distributions of different dates is found, alleviating the problem of low estimation accuracy caused by small sample size; different transformations can be applied to a non-stationary time series to make it close to stationary; because there are several types of stationarity, we can use single root test methods, commonly used test methods are ADF test (Augmented Dicky Fuller Test) and KPSS test (Kwiatkowski-Phillips-Schmidt-Shin Test): If the ADF test result is stationary and the KPSS test result is non-stationary, the time series is difference stationary - apply differences to the time series and check stationarity again; if the ADF test result is non-stationary and the KPSS test result is stationary, the time series is trend stationary, which requires deterministic detrending and stationarity check again.

[0042] In a possible implementation, in step S2, constructing a city network according to the geographical distribution map of the region and its neighboring information specifically includes: first, according to the definition of the geographical distribution map, the regional network of the monitored area is represented as an undirected graph G = (V, E, W), where each node represents a region, and each edge represents the adjacency relationship between two regions;

[0043] represents a set of M vertices, represents a set of N edges in the network, and the relationship between each edge can be represented by two vertices (vi, vj); W represents an M*M spatial weight matrix, that is, vertices vi and vj are adjacent, then W ijis assigned a value of 1, otherwise it is assigned a value of 0; the epidemic-related mixed time series data of each region are synchronized to form multiple groups of time series data with the structure of St = {s1, s2, ..., st}; the data is combined with the regional network to construct a set of dynamic infectious disease regional network diagrams that change over time. In specific applications, the present invention intends to use the epidemic mixed data prediction model of the time series attribute-attention mechanism to predict the optimized data after the data optimization operation, so as to realize the dynamic prediction of multi-source epidemic data. In terms of predictive modeling research in the spatial dimension, the statistical space theory is introduced into the spatial mixing prediction model, and the multidimensional spatial mixing prediction model based on spatial weights maps the complex relationship between the mixed data variables and the spatial dependency of the variables into the form of an adjacency matrix.

[0044] In one possible implementation, the dynamic infectious disease regional network includes spatiotemporal data processing, multi-channel fusion learning, and network enhancement connection strategy. In specific applications, the method of aggregating neighbor node information based on graph neural networks is used to fuse mixed data, thereby achieving the purpose of mixed data representation learning; this article needs to explore different optimization algorithms, adjust network structure and parameters, and perform cross-validation to evaluate model performance; optimize deep learning models to improve prediction accuracy and efficiency; and explore the application of different deep learning models in epidemic data analysis and prediction; and evaluate the effectiveness of these models in processing epidemic time series data (such as infection rate, transmission speed), including model training, verification and testing processes; reduce prediction errors. In addition, considering the time-varying nature of epidemic data, it is necessary to improve the real-time prediction of future epidemics; the present invention will also focus on the performance of the model in processing time series data, thereby significantly improving the accuracy and response speed of the prediction.

[0045] The basic principles of the present invention are as follows: (1) constructing an adaptive, high-precision multivariate epidemic prediction model, using the accuracy and reliability of the traditional ARIMA model on linear time series data, combined with the powerful fitting ability of deep learning on nonlinear data, while ensuring the stability and applicability of the model, further improving its prediction accuracy, thereby achieving the accuracy of epidemic data prediction. The present invention uses deep learning technology to improve the accuracy of epidemic prediction, including predictions of disease development, such as infection rate, transmission speed and geographical distribution;

[0046] (2) Construct an epidemic prediction model that meets the needs of mixed multidimensional data. The present invention intends to solve the optimization problem of spatial mixed prediction models in a multi-source heterogeneous big data environment. Starting from the characteristics of mixed data, a new method for early warning and monitoring of heterogeneous multidimensional mixed data is established from the time series characteristics, spatial characteristics and data frequency characteristics. The present invention is also committed to improving the interpretability of the model so that operators can understand why the model makes specific predictions, thereby improving the credibility and application value of the model, enabling the model to be applied to the actual research of epidemic data, and proposing an application demonstration prototype.

[0047] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting epidemic mixed frequency data based on deep learning, characterized in that: The following steps are involved: S1. Obtain historical epidemic data, real-time monitoring data, and demographic data, and associate data of different frequencies to obtain mixed frequency data through information fusion; S2, determining the size of the sliding window according to the optimal lag order and sampling time interval of the mixing data, and determining the data length of the data of each sampling frequency in the window when the sliding window is initialized; constructing a city network according to the geographical distribution map of the region and its adjacent information; S3, learn the time-dependent features of each frequency data through the LSTM network and obtain the mixed frequency data embedding vector; S4, adaptively extracting or expanding the timing characteristics in the time dimension through a CNN network with the same number of convolution kernels, wherein the CNN network unifies the mixed data into the same data length and fuses them to obtain fusion features, so that the feature layer information of the mixed data is fused; S5. Learning an attention vector from the fused feature through a time attribute attention mechanism, and obtaining a time and attribute feature context vector, wherein the attention mechanism adaptively mines important features in the fused feature; S6. The mixed data embedding vector, time and attribute feature context vector are output through a fully connected layer to obtain a prediction value.

2. The method for predicting epidemic mixed frequency data based on deep learning according to claim 1, characterized in that: The temporal attribute attention mechanism is a two-dimensional attention mechanism, which extracts information from two dimensions, time and attribute, respectively, so that the temporal attribute attention mechanism generates two context-related vectors.

3. The method for predicting epidemic mixed frequency data based on deep learning according to claim 1, characterized in that: The mixed frequency data needs to be processed, and the data processing includes data preprocessing and data sequence stabilization.

4. The method for predicting epidemic mixed frequency data based on deep learning according to claim 3, characterized in that: The data preprocessing includes data cleaning, missing value processing and outlier detection.

5. The method for predicting epidemic mixed frequency data based on deep learning according to claim 3, characterized in that: The data series smoothing includes differencing and deterministic detrending.

6. The method for predicting epidemic mixed frequency data based on deep learning according to claim 5, characterized in that: The data series is tested for stationarity by ADF and KPSS. When the ADF test data series result is stationary and the KPSS test data series result is non-stationary, the time series is differentially stationary - the time series is differentially applied and the stationarity is checked again. When the ADF test result is non-stationary and the KPSS test result is stationary, the time series is trend-stationary and requires deterministic detrending and stationarity is checked again.

7. The method for predicting epidemic mixed frequency data based on deep learning according to claim 1 is characterized in that: step In step S2, the construction of the urban network according to the geographical distribution map of the region and its adjacent information specifically includes, firstly, according to the definition of the geographical distribution map, the regional network of the monitored area is represented as an undirected graph G = (V, E, W), wherein each node represents a region, and each edge represents the adjacency relationship between two regions; represents a set of M vertices, represents a set of N edges in the network, and the relationship between each edge can be represented by two vertices (vi, vj); W represents an M*M spatial weight matrix, that is, vertices vi and vj are adjacent, then W ij is assigned a value of 1, otherwise it is assigned a value of 0; the mixed-frequency time series data related to the epidemic in each region are synchronized to form multiple sets of homofrequency time series data, whose structure is St = {s1, s2, ..., st}; the data are combined with the regional network to construct a set of dynamic infectious disease regional network diagrams that change over time.

8. The method for predicting epidemic mixed frequency data based on deep learning according to claim 7, characterized in that: The dynamic infectious disease regional network includes spatiotemporal data processing, multi-channel fusion learning and network enhanced connection strategy.

Images